AI-driven Grid Optimization Can Reduce Emissions (Papers Track)
Kyri Baker (Google DeepMind); Jackie Kay (Google DeepMind); Miha Zgubic (Google DeepMind); Eric Perim Martins (Google DeepMind); Sofia Liguori (Google DeepMind); Steven Bohez (Google DeepMind); Sephora Madjiheurem (Google DeepMind); Marc Deisenroth (Google DeepMind); Laura Toni (Google DeepMind); Sims Witherspoon (Google DeepMind); Sophie Elster (Google DeepMind); Kyle Levin (Google DeepMind); Luis Piloto (Google DeepMind)
Abstract
Power systems are an essential backbone of modern society, and notoriously hard to operate optimally, as supply and demand must be balanced in near real-time. Due to the nonlinear dynamics of these networks, many grid operators in the United States and worldwide use linear approximations to clear markets and optimize generator setpoints, introducing inefficiencies like increased losses and unnecessary excess generation. In this paper, we show that the carbon footprint incurred by training a model to learn AC optimal power flow solutions is drastically offset by the gains in operational efficiency (in terms of wasted energy generation) from using these models to optimize grid operations. In particular, we show that it generally takes on the timescale of minutes for these models to offset their initial training footprint.